OpenAI’s decision to retire GPT-4o from ChatGPT on February 13, 2026 has set off an unusually emotional response from a segment of its users, reviving a debate about how people form attachments to conversational AI and where companies should draw the line.
The company announced the retirement on January 29, saying GPT-4o and several older models would be removed from ChatGPT in favor of its newer GPT-5 series. In the days that followed, users on social platforms and forums described the model as a friend, a therapist and a source of emotional support rather than a chatbot. Some wrote that it had become part of their daily routine; others said it had helped them stay mentally balanced. Petitions circulated asking OpenAI to keep the model available or, failing that, to release its weights publicly, and at least one open letter addressed to chief executive Sam Altman described the system as a source of calm and stability.
A small share, but a large number
OpenAI has said that only a very small fraction of ChatGPT users engaged in deep emotional conversations with GPT-4o, and that a similarly small share still chose the model daily by the time of the announcement. With a user base in the hundreds of millions, however, even a fraction of a percent represents a large number of people.
This is not the first time the company has faced this kind of pressure. When GPT-5 launched in August 2025 and GPT-4o was initially removed, the protest from users was strong enough that OpenAI restored the older model for paying subscribers within days. The February retirement is the second attempt to phase it out.
Lawsuits over the model’s warmth
The attachment some users feel has a darker counterpart. OpenAI faces a series of lawsuits, including at least seven filed in California state courts in November 2025, alleging that GPT-4o’s highly empathetic and validating responses deepened mental health crises in some users. The complaints argue that the model fostered unhealthy dependence and that safeguards such as redirecting users to real-world help or ending risky conversations were not applied effectively. OpenAI has disputed the characterizations and said it is reviewing the filings.
The company’s position is that ChatGPT is trained to recognize signs of distress, de-escalate conversations and point users toward professional help. It says newer models include stricter safety behaviors intended to prevent unhealthy reliance on the system. For some users, though, those same guardrails are precisely what they dislike: they experience the newer models as colder and less personal than the one being retired.
The broader question
The GPT-4o episode illustrates a tension that is likely to recur as conversational AI becomes more capable. The same qualities that make a model comforting to someone who is lonely or under stress can, in other circumstances, reinforce harmful patterns. Regulators, clinicians and the companies themselves have yet to settle on where the balance lies, and the reaction to this retirement suggests that decisions about model behavior now carry consequences well beyond product design.
